better than baseline link prediction on the patent ... · jaccard observation: 1. worst asymptotic...
TRANSCRIPT
Link Prediction on the Patent Citation Network
Problem Definition
Experimental Results
Motivation
Datasets
Our Approach
○○
○○
Approach 1: SDNE
Approach 3: Jaccard Coefficient
○
References1. Grover and Leskovec. node2vec: Scalable Feature Learning for Networks. ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (KDD), 2016.2. Rodriguez et al. Graph mining algorithms for the analysis of patent citation networks, Dissertation, Rutgers
University.3. Wang et al. Structural deep network embedding. In Proceedings of the 22nd ACM SIGKDD international
conference on Knowledge discovery and data mining. ACM 2016, 1225–1234.4. Niwattanakul et al. Using of Jaccard coefficient for keywords similarity. In Proceedings of the international
multiconference of engineers and computer scientists 2013 (Vol. 1, No. 6, pp. 380-384).5. Lerer et al. (2019). PyTorch-BigGraph: A Large-scale Graph Embedding System. arXiv preprint 2019,
arXiv:1903.12287.
Conclusions
1. node2vec: Much more computationally efficient than SDNE
2. node2vec: Performs better than SDNE, particularly on sparse graph
3. node2vec: Prediction on future is better than baseline
Jaccard Observation:1. Worst asymptotic complexity
○
•
○••
EECS 598-008, W19Advanced Data
Mining
•○○○
Blog Catalog
1988-1989 Baseline
1988-1999 Future
•••➢
Approach 2: node2vec
•○○○○
Data Partitioning
Link Prediction from Node Embeddings
○
Approach 4: PyTorch BigGraph
BigGraph Observation:1. Performance increases with k.2. Highly scalable